AI codes better than me. Now what?

Context

Lee Robinson, now at Cursor after years at Vercel, built a Rust-based image compressor, a SvelteKit web app, and a hardware game for the Analog Pocket—all without writing code by hand. He used coding agents exclusively.

Key Takeaways

Models will get good at everything. Don't assume your niche is safe. Rust, low-level code, computer science algorithms—models handle them now. Within a year, expect competence across most software engineering tasks.

Eliminate drudge work first. Shell scripts, JSON wrangling, cryptic error messages—offload the boring stuff. Coding agents excel at boilerplate. Focus your energy on what matters for the product.

Become a generalist. Code is cheap now. Taste matters more. Product engineers who blend coding with UX, marketing, and user research will thrive. Build something end-to-end: ship it, market it, grow it.

Learn deliberately. LLMs can trick you into thinking you understand something. Don't delegate your thinking. Ask the model to generate mini-courses, explain concepts layer by layer. Use it as an on-demand tutor, not a crutch.

Notable Quotes

"Writing code was never really the bottleneck, especially for larger projects."

"It wasn't about the code... It's about building something great and something that I'm proud of."

Connections

This video captures the same shift described in 12-factor-agents and building-effective-agents—the move from writing code to orchestrating agents. For practical tips on working with these tools, see claude-code-best-practices and how-i-use-llms.